Taxonomy, Descriptions, and Individuals in Natural Language Understanding

Ronald J. BrachmanACL 1979 · DOI

Introduction

This foundational 1979 paper by Ronald J. Brachman introduces a prototype natural language understanding (NLU) system built on KL-ONE, a sophisticated knowledge representation (KR) language. At a time when AI was grappling with how to bridge the gap between raw syntax and deep meaning, Brachman proposed using structured inheritance networks to handle the "messiness" of human dialogue—including indirect speech acts, anaphora, and complex quantification.

The system leverages a formal taxonomy to resolve ambiguous references and infer a user's true intent (e.g., recognizing that "I want to see X" is a command to the display) without prematurely committing to a specific logical scope.

Problem Statement: Bridging the Gap in NLU

In the late 1970s, NLU systems often struggled with three major hurdles:

  1. Ambiguity and Scope: Representing sentences like "Show me the three states" without immediately deciding exactly which three states or how they relate to the display.
  2. Indirect Speech Acts: Humans rarely say exactly what they mean. "I want to see the network" is a statement of desire, but in a computer interface, it's a command.
  3. Knowledge Integration: There was a disconnect between the "grammar" of a language and the "knowledge" of the world.

Brachman sought a way to represent "quantificational import"—the meaning of sets and numbers in a sentence—while keeping the representation flexible enough for a discourse expert to refine later.

Methodology and System Architecture

The core of the system is KL-ONE, a language based on "structured inheritance networks." The architecture integrates syntactic parsing, semantic interpretation, and discourse reasoning into a unified flow.

Figure 1: Overview of the NLU system architecture, showing the integration of the RUS parser, discourse expert, and KL-ONE knowledge bases.

Knowledge Representation in KL-ONE

  • Taxonomic Inheritance: Knowledge is organized into a hierarchy of Concepts (generic templates) and Roles (properties or parts). Sub-concepts inherit properties from their parents but can add specific restrictions.
  • Automatic Concept Descent: The system’s parser (a version of the RUS system) doesn't just label parts of speech; it uses the KL-ONE taxonomy as a "discrimination net." As it gathers more information about a phrase, it "pushes" the concept down the hierarchy to the most specific possible definition.
  • DSETs (Determined Sets): To handle plurals and quantifiers, the authors introduced DSETs. These allow the system to represent a "set of entities" as a single unit, capturing the intent of a plural noun phrase without needing to enumerate every member immediately.
Figure 4: KL-ONE representation of a Determined Set (DSET), illustrating how the system handles quantificational import and plural noun phrases.
  • Discourse Expert: This component uses plan-recognition to bridge the gap between "surface" speech acts (what was said) and "intended" speech acts (what the user wants done).

Implementation Domain

The system was implemented as a natural language interface for a specific technical domain: the Augmented Transition Network (ATN) grammar of the LUNAR system. The "world" the system knows about consists of the states, arcs, and conditions of this specific grammar. Evaluation was qualitative, based on the system's ability to handle a sample dialogue involving commands to display, highlight, and "focus in" on parts of the ATN.

Key Insights

  • Syntax Guided by Semantics: By embedding the dictionary in a syntactic taxonomy, the parser can rule out nonsensical paths early. For example, if a "LOCATION-PP" requires a "PLACE-NOUN," the parser won't even try to build a location phrase using a word it knows isn't a place.
  • Definitional vs. Assertional: KL-ONE distinguishes between what a thing is by definition (e.g., a "Connecting-Arc" must have a "nextState") and what happens to be true in a specific context.
  • The "Backing Off" Maneuver: The system handles corrections gracefully. If a user says "No, I want to see X," the discourse expert recognizes this as a request to modify the current display state rather than starting a new one from scratch.
  • Procedural Attachment: KL-ONE allows "hooks" to be attached to concepts. When a concept is "individuated" (created as an instance), it can automatically trigger procedures—like drawing a shape on the screen.

Results and Discussion

The prototype demonstrated sophisticated interaction capabilities:

  • Anaphora Resolution: The system could track "the states" mentioned in previous sentences.
  • Intent Inference: It successfully mapped surface statements (e.g., "I want to be able to see S/AUX") to system actions (e.g., "INCLUDE S/AUX in the display").
  • Structural Precision: The use of KL-ONE allowed the system to represent the literal semantic content of an utterance exactly as specified, without "over-committing" to interpretations that weren't overtly in the text.

Limitations and Future Work

Several features, such as the "language generation" component and the "Nexus" (a construct for handling coreference across different contexts), were proposed but not yet fully implemented at the time of publication. Additionally, while the KL-ONE framework is general-purpose, the prototype was tightly coupled to the LUNAR ATN domain. Future work would need to address the computational overhead of maintaining and traversing very large inheritance networks, a common challenge for frame-based systems.

Summary updated

BibTeX
@inproceedings{brachman1979taxonomy,
  author = {Ronald J. Brachman},
  title = {{Taxonomy, Descriptions, and Individuals in Natural Language Understanding}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1979},
  doi = {10.3115/982163.982174},
}
Figure